\singlespacing

\newpage


\setcounter{table}{0}
\renewcommand{\thetable}{A\arabic{table}} 
\setcounter{figure}{0} 
\renewcommand{\thefigure}{A\arabic{figure}}

\setcounter{page}{1}

# Online Appendix A: Formalization of the Argument {#gamemodel -}

The formalization of the affirmation propaganda argument outlined here is adapted from theoretical models of Bayesian persuasion [@Kamenica.Gentzkow2011], including their application to media control [@GehlbachSonin2014]. In these models, one actor, the sender, aims to persuade another actor, the receiver, to take an action that the sender prefers rather than the action that the receiver prefers in the absence of the sender's messages. The formalization here incorporates heterogeneity of prior beliefs among receivers, which in this context corresponds to pro-regime or oppositional attitudes.^[I am grateful to Scott Gehlbach for suggesting this approach to the formalization.] The analysis below demonstrates that under certain conditions, autocrats have to choose between maintaining existing support and convincing the unpersuaded.

The autocrat is the sender, and the citizens are the receivers. There are two groups of citizens, $A$ (the pro-regime majority) and $B$ (the opposition, or the minority), of sizes $\alpha_{A}$ and $\alpha_{B}$, where $\alpha_{A} > \alpha_{B}$, and $\alpha_{A} + \alpha_{B} = 1$. 

The state of the world $\theta \in \{0, 1\}$ is a random variable, unobserved by autocrat and citizens. The variable $\theta$ may represent, e.g., economic or government performance; $\theta=1$ means that the state of the world is good. Citizens do not observe the state of the world, and they must choose an action $a \in \{a_{0}, a_{1}\}$, e.g., $a_{1}$ could be voting for the autocrat, and $a_{0}$ would be voting against. Citizens' payoffs are dependent on their action and on the state of the world: for any citizen $i$, the payoff is $x$ if $\theta = 0$ and $a_{i} = a_{0}$, $1-x$ if $\theta = 1$ and $a_{i} = a_{1}$, and 0 otherwise.

In a departure from the standard framework, I assume that citizens have heterogeneous prior beliefs about the state of the world, $p_{A} > x$ and $p_{B} < x$, where $p_B$ is the weight group $B$ places on the event $\theta = 1$. That is, group $A$ is ex ante inclined to take the autocrat's preferred action $a_{1}$, and group $B$ is ex ante not inclined to take that action. The autocrat's payoff is equal to the share of citizens that take the action $a_{1}$.

Before the state of the world is realized, the autocrat commits to a "signal structure," which is a probability distribution over messages for each state of the world. With probability $\beta_{\theta}$, the autocrat sends the propaganda message $m = 1$. Without loss of generality, I assume $\beta_{1} = 1$, so that the news is always "good" when the state of the world is "good." Of primary interest is $\beta_{0}$, which can be interpreted as media bias. 

The state of the world is then realized, and the propaganda message is generated based on $\beta$. Citizens then update their beliefs using Bayes' rule and choose the action $a$.

What is the level of media bias $\beta_{0}$ that maximizes the autocrat's payoff? The choice of $\beta_{0}$ by the autocrat is constrained by the conditions under which the receivers would take the sender's preferred action when $m = 1$; following @BergemannMorris2019, I refer to these conditions as obedience constraints. I ask: If there are two groups of citizens with different priors, when is it optimal for the autocrat to set media bias $\beta_{0}$ such that the obedience constraint for group $B$ is satisfied ($B$ takes the action $a_{1}$), and when is it, instead, optimal to simply focus on satisfying the constraint for group $A$ (ensuring that $A$ is still willing to take the action)?

It is always possible to ensure that group $A$ (the majority) takes the autocrat's preferred action as long as the autocrat is willing to forgo persuading group $B$ (the opposition). For example, if the autocrat sets $\beta_{0} = 1$, propaganda always sends a positive signal ($m = 1$), and there is no updating for either group. The autocrat's expected payoff in this case is $\alpha_A$ (the share of $A$ in the population), as only citizens in $A$ choose $a_{1}$. 

However, the reverse is not true: if the autocrat persuades group $B$ to take the action, it is possible that group $A$ will not take the action. To satisfy the obedience constraint for $B$, media bias $\beta_{0}$ should be sufficiently low so that $m = 1$ could be an informative message for $B$. Given that $Pr_{B}(\theta = 1) = p_{B}$, the obedience constraint for $B$ is $\frac{p_{B}}{p_{B} + (1-p_{B})*\beta_{0}} \geq x$. Rearranging, media bias such that the obedience constraint binds for $B$ is $\beta_{0} = \frac{p_{B}}{1-p_{B}}*\frac{1-x}{x}$.

Implementing media bias to convince $B$ means that sometimes the autocrat must send $m=0$ when $\theta=0$. When this is the case, group $A$ (the majority) will also infer that $\theta = 0$ and not take the action preferred by the autocrat. 

The choice between two strategies---targeting only the majority versus attempting also to persuade the opposition---depends on the various parameters of the model. As shown above, the payoff from the first strategy is $\alpha_{A}$. To define the autocrat's expected payoff in the second case, posit an (ad hoc) "true" prior $p = Pr(\theta = 1)$. Then, the autocrat's expected payoff is $p + (1-p)*\frac{p_{B}}{1-p_{B}}*\frac{1-x}{x}$, given the optimal media bias derived above. 

The autocrat thus focuses on convincing $B$ if $p + (1-p)*\frac{p_{B}}{1-p_{B}}*\frac{1-x}{x} > \alpha_A$, so the choice depends on the size of the majority ($\alpha_A$) and on $p_{B}$. Reaching out "across the aisle" can be beneficial only if $p_{B}$ is sufficiently large (close to $x$), so the autocrat can win $B$ over by sending $m=0$ only occasionally, and if $\alpha_A$ is relatively small.

With small values of $p_{B}$---if $p_{B}$ is distant from $x$ and, therefore, from $p_{A}$---autocrats need to send informative messages ($m=0$) often if they want to win over the highly skeptical opposition, but such messages would also alienate many members of the majority. In other words, if there is a large divergence in priors between the supporting majority and the opposition, it is not optimal for the autocrat to cater to the latter. Further, if the size of the ex-ante pro-regime group is large enough, the autocrat can simply produce uninformative (positive) messages all of the time regardless of the difference in priors between the two groups.

The situation when there is a strong majority that supports the autocrat and the opposition is small but ideologically distant is observed in certain authoritarian regimes. In this environment, the autocrat would in equilibrium choose substantial media bias that targets the majority group alone---that is, would choose affirmation propaganda.

\newpage

\setcounter{table}{0}
\renewcommand{\thetable}{B\arabic{table}} 
\setcounter{figure}{0} 
\renewcommand{\thefigure}{B\arabic{figure}}

\setcounter{page}{1}

# Online Appendix B: Additional Evidence From the Main Study (the Social Media Sample) {#appendixB -}

## A Note on Human Subjects Research {-}

This study was determined to be exempt by the Institutional Review Board at the University of Wisconsin-Madison (IRB protocols ID 2019-0763, 2019-0800, and 2020-0639), as defined under 45 CFR 46 (Category 2). For questions, you may contact the Education and Social/Behavioral Science IRB at 608-263-2320. The study is in compliance with APSA’s Principles and Guidance for Human Subjects Research. In particular, the participants were Russian adults who engaged with the study using their native language; the participants provided their informed consent to participate in the study; the study did not collect any identifying data on the participants; their responses are kept confidential and are analyzed only in an aggregated form. The sample size was determined based on the number of experimental treatments and the heterogeneous effects that were to be examined. 

The experiment on the social media sample and the survey experiment embedded in the Levada survey involved slight deception---specifically, some participants might have seen news messages attributed to news sources that had not actually published these news stories, and the purpose of the study was not fully disclosed in the beginning of the surveys. In both cases, the deception was necessary in order to avoid demand effects and other distortions: if participants were aware that the purpose of the study was to understand their news source perceptions and the relationship between source perceptions and political views, they might not have answered truthfully. The purpose of the study and the nature of the experimental manipulation were fully disclosed to participants in the debriefing message displayed after the completion of each survey. The subjects were able to contact the researcher in case they had any questions. 

```{r vignette, echo = FALSE, fig.show = "hold", out.width = "50%", fig.align = "center", fig.cap = "This is an example of an experimental vignette with a news story attributed to a state-controlled news outlet, Russia-24"}
knitr::include_graphics("exp_vignette_sample_eng_pope.png")
```


\newpage

## Summary Statistics {-}

```{r sumstats3, echo = FALSE, results = "asis"}

sumstats_resp <- read_csv("Estimates/Table_B1_sumstats_resp_3_surveys.csv")

kable(sumstats_resp, format = "latex",
      caption = "Summary statistics for the three samples", booktabs = T, linesep = "",
      align = c("l", "c", "c", "c", "c", "c", "c"),
      col.names = c("Variable", "%", "Non-missing", "%", "Non-missing", "%", "Non-missing")) %>%
  add_header_above(c(" ", "Main study" = 2, "National survey" = 2, "Online panel" = 2)) %>%
  kable_styling(latex_options = c("hold_position"), font_size = 9) %>%
  column_spec(1, width = "14em") %>%
  column_spec(c(2, 4, 6), width = "3em") %>%
  column_spec(c(3, 5, 7), width = "6em") %>%
  footnote(general = "The sample is limited to respondents with non-missing data on presidential approval.",
           threeparttable = T, footnote_as_chunk = T)

```

\newpage


## The Procedure for the Selection of News Stories {-}

Fourteen news stories in the main quiz and 16 stories in the second quiz (see the main text for details) were selected from top news stories by Russian online news aggregators in the months preceding the study. Several news stories were sought and included specifically to ensure, first, that there were some false news stories in the list, and second, that there were pro-Russia, critical, and neutral stories.

To check the veracity of these news stories, I relied on existing fact-checking resources such as _Politifact_ and the fact checks regularly published by the Russian investigative web site _The Insider_. When existing fact checks were not available, I fact checked the stories based on reports by authoritative independent news agencies, economic reports, and other data. If the veracity of a story could not be established, the story was excluded from selection.

Two slots in the quiz were reserved for "current" stories that were updated two or three times a week based on recent news reports. First, I used a web scraping script to download top news stories on politics and international news from _Yandex News_, Russia's largest news aggregator with a daily audience of 9 million people. _Yandex_ uses an algorithm to determine the news stories that are popular at any given moment. "Politics" and "world news" are two of the sections on the _Yandex News_ main page, and at any particular moment, there are several dozens of news stories under each of these two labels. 

After downloading all the stories in these two categories, I eliminated irrelevant messages based on several criteria: stories that reported future events without indicating their substance (e.g., announcements of press conferences); stories that were developing and could change quickly (e.g., the number of deaths from COVID-19); stories focused on technical details of events (e.g., the amount of shipments entering a port, low-level bureaucratic appointments); opinions or personal statements, except for statements by key political and business leaders; stories that could not be reliably fact-checked (e.g., information about military operations).

This preliminary selection produced shorter lists of candidate news stories under both "politics" and "world news." After obtaining these lists, I used a random number generator to select one news story from each of the two topics. These two news stories were fact-checked and then added to the survey. Largely, I aimed to preserve the headlines from _Yandex News_, sometimes expanding the headline based on the text of the corresponding news story or slightly editing it for clarity. 

\newpage

## The Categorization of State-Controlled and Independent Media Outlets {-}

Various analyses in this study rely on a categorization of news outlets as state-controlled or independent. This subsection lists all the news outlets that are used in the study either as experimental treatments or as answer choices in questions about media trust and media usage. News outlets that are included as treatments in the experiment are in **bold**.

**State-controlled media outlets**: _**Channel One**_, _**Russia-24**_, _Russia-1_, _Vesti_, _**RT**_, _**RIA**_, _TASS_, _Zvezda_, _Sputnik_, _Rossiyskaya Gazeta (RG)_ (all of the preceding outlets are owned by the government); _NTV_, _RenTV_, _**Komsomolskaya Pravda (KP)**_, _Moskovskiy Komsomolets_, _Izvestiya_, _Lenta.ru_, _Gazeta.ru_, _Vzglyad_ (these outlets were controlled by pro-Kremlin oligarchs).

**Independent (critical) media outlets**: _**Rain**_, _Novaya Gazeta_, _Vedomosti_, _Rosbalt_ (owned by independent entrepreneurs); _**Echo of Moscow**_; _BBC_, _**Meduza**_, _Euronews_, and other foreign news sources.

The list of news outlets also included _RBC_ and _Kommersant_, business news outlets that were controlled by Kremlin-friendly oligarchs but were not as propagandistic as the state-controlled media organizations listed above.

This list of news outlets was compiled based on several internet rankings of most popular websites in Russia (_Yandex.Radar_, _Liveinternet_, _Rambler Top 100_, _Mediametrics_), and some less popular, but important independent news outlets such as _BBC_ were added. 

The categorization into state-controlled and critical news outlets is based on media ownership, on news reports on the Russian media industry, and on previous scholarship that has examined or categorized Russian media [@SimonovRao2022; @GreeneRobertson2019; @Schimpfoessl.Yablokov2017].

\newpage


## News Stories in the Experiment {-}

```{r stories2020, echo = FALSE}
stories_data <- read_csv("Estimates/Table_B2_story_summary_2020.csv")

kable(stories_data, format = "latex",
        caption = "News messages evaluated in the main study", longtable = T, booktabs = T, linesep = "",
        align = c("c", "l", "c", "c", "c", "c", "c", "c")) %>%
    kable_styling(latex_options = c("hold_position"), font_size = 9) %>%
    column_spec(1, width = "2em") %>%
    column_spec(2, width = "15em") %>%
    column_spec(3, width = "4em") %>%
    column_spec(4, width = "5em") %>%
    column_spec(5, width = "6em") %>%
    column_spec(6, width = "4em") %>%
    column_spec(7, width = "4em") %>%
    column_spec(8, width = "5em")%>%
    add_header_above(c(" " = 5, "Mean evaluations" = 3)) %>%
    footnote(general = "The last three columns present the proportion of those who evaluated the corresponding story as true in the full sample, among Putin supporters, and among Putin critics, respectively. Stories 1-30 are 'pre-selected,' and stories 31-50 are 'current.' Stories 1-14 and 31-50 included in the first quiz, stories 15-30 included in the second quiz. See the text for details. Story 3 was also included in the nationally representative survey (Study 2). Stories 7, 10, and 11 were also included in the OMI online panel (Study 3).",
  threeparttable = T, footnote_as_chunk = T)


```

\newpage

## Putin Supporters Are More Receptive to Propaganda {-}

```{r storyperc2, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=5, out.width = "85%", fig.align = "center", fig.showtext = TRUE, fig.cap = "Covariate-adjusted differences in the shares of respondents who found stories credible. Calculated from linear regressions of story evaluations on Putin approval and covariates. Results from the main study. 95\\% confidence intervals are shown."}

stories_perc2 <- read_csv("Estimates/Figure_2_B2_stories_perc_by_approval_adj.csv") %>%
filter(model == "Average") %>%
mutate(Stories = str_remove(Stories, " stories, mean"))

ggplot(stories_perc2, 
       aes(Stories, diff)) +
  geom_point(size = 3,
             position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 1) +
  geom_text(aes(label = round(diff, 1)), vjust = -0.5, size = 4.5, 
            family = "sourcesanspro", show.legend = F) +
  geom_hline(yintercept = 0, linetype = "dashed", col = "gray") +
  labs(x = "", y = "Difference, percentage points",
       title = "") +
#  ylim(c(-0.27, 0.27)) +
  coord_flip() +
  myggtheme +
  facet_wrap(~contrast, ncol = 3)


```

Figure \@ref(fig:storiesthreesurveys) compares the differences between Putin critics and supporters in evaluations of selected stories between the main study and the two additional surveys. The story labels refer to the following stories in Table \@ref(tab:stories2020): "Growth in Ukraine"---story 3; "Trump and Putin"---story 7; "COVID origins"---story 10; "Nuclear waste"---story 11.

```{r storiesthreesurveys, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=7, out.width = "85%", fig.align = "center", fig.showtext=T, fig.cap = "Covariate-adjusted differences in the shares of Putin supporters and critics who found stories credible. Results from Studies 1, 2, and 3. 95\\% confidence intervals are shown"}

all_contrasts_approval <- read_csv("Estimates/Figure_B3_stories_perc_by_approval_3_surveys.csv")

ggplot(all_contrasts_approval, 
       aes(story_label, estimate, linetype = Survey, shape = Survey)) +
  geom_point(size = 3,
             position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 0.8) +
  geom_hline(yintercept = 0, linetype = "dashed", col = "gray") +
  labs(x = "", y = "Difference between\nsupporters and critics",
       title = "Supporters and critics disagree about news stories in all samples") +
  ylim(c(-30, 27)) +
  myggtheme +
  facet_wrap(~story_direction, ncol = 2, scales = "free_x")
```


\newpage

## Balance Check {-}

```{r balanceonline, echo = FALSE}

balance_o <- readRDS("Estimates/Table_B3_balance_check_main_study.rds")

kable(balance_o, format = "latex",
        caption = "Covariate balance check for the experiment (the main study)", booktabs = T, linesep = "",
        align = c("l", "c", "c", "c", "c", "c", "c", "c")) %>%
    kable_styling(latex_options = c("hold_position"), font_size = 9) %>%
    column_spec(1, width = "6em") %>%
    column_spec(2, width = "5em") %>%
    column_spec(3, width = "5em") %>%
    column_spec(4, width = "5em") %>%
    column_spec(5, width = "5em") %>%
    column_spec(6, width = "5em") %>%
    column_spec(7, width = "5em") %>%
    column_spec(8, width = "5em") %>%
    footnote(general = "Results of chi-square test for equality of covariate values across treatment groups, by news story. In each column, I provide p-values from chi-squared tests of equality of covariate values across treatment groups (news sources) for the corresponding covariate. See story texts in the list of stories above.",
  threeparttable = T, footnote_as_chunk = T)

```


\newpage

## Experimental Results with Other Measures of Pro-Regime Orientations {-} 

As discussed in the main text, empirical evidence suggests that Russians are generally truthful when reporting their presidential approval. Nonetheless, I have implemented additional measures to improve the robustness of results. First, I asked the respondents about events or developments in Russian history they are proud of. One of the possible answers was "the reunion with Crimea" (the annexation of Crimea in 2014), very popular among Putin supporters but not among critics. The correlation between presidential approval and pride in the annexation was about 0.48.

Second, in the beginning of the quiz, respondents evaluated two news stories. One reported that the European Union had lost 500 billion euros because of sanctions against Russia (an untrue propaganda statement spread by Vladimir Putin). The other story reported that the Ukrainian economy had been growing faster than the Russian economy (a true story incongruent with common beliefs of government loyalists, as Ukraine was typically portrayed in Russian state media as a failed state). In the quiz, these stories were always attributed to one news source, a news agency _Interfax_.

Then, I combined responses to these two statements in an index that takes the value of 2 if a respondent finds the pro-government EU story to be true and the Ukraine story to be false, the value of 0 if a respondent finds the EU story to be false and the Ukraine story to be true, and the value of 1 if both stories are found to be false or both are found to be true. Larger values are consistent with stronger pro-regime sympathies. The correlation between presidential approval and this measure is about 0.32.

Figure \@ref(fig:onlinestateeffectpride) shows the effect of switching from critical to state media depending on pride in Crimea and on beliefs about EU and Ukraine; regression models are in Table \@ref(tabonlinestatemediaeffectaltapproval). 

```{r onlinestateeffectpride, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=6, out.width = "85%", fig.align = "center", fig.showtext=T, fig.cap = "The effect of changing the treatment from critical to state media outlet on evaluations of news stories. Calculations based on a linear regression of news story evaluations, accounting for state control and pro-regime orientations; results from the main study. 95\\% confidence intervals are shown"}

library(patchwork)

pride <- read_csv("Estimates/Figure_B4_main_treatment_effect_pride_crimea.csv")
eu <- read_csv("Estimates/Figure_B4_main_treatment_effect_eu_ukr.csv")

gp <- ggplot(pride, 
       aes(pride_history_crimea_cat, estimate)) +
  geom_point(position = position_dodge(width = 0.4), size = 3) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 0.8) +
  geom_text(aes(label = round(estimate, 1)), vjust = -0.5, size = 4.5, 
            family = "sourcesanspro", show.legend = F) +
  geom_hline(yintercept = 0, linetype = "dashed", col = "gray") +
  labs(x = "", y = "Effect of state media",
       title = "Given feelings toward Crimea") +
  ylim(c(-17, 17)) +
  myggtheme +
  coord_flip()

ge <- ggplot(eu, 
       aes(EU_Ukr_beliefs, estimate)) +
  geom_point(position = position_dodge(width = 0.4), size = 3) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 0.8) +
  geom_text(aes(label = round(estimate, 1)), vjust = -0.5, size = 4.5, 
            family = "sourcesanspro", show.legend = F) +
  geom_hline(yintercept = 0, linetype = "dashed", col = "gray") +
  labs(x = "", y = "Effect of state media",
       title = "Given beliefs about EU/Ukraine") +
  ylim(c(-17, 17)) +
  myggtheme +
  coord_flip()

gp + ge + plot_annotation(title = "Effect of state vs critical media", theme = theme(plot.title = element_text(size = 30,
                                    family = "sourcesanspro")))


```

\newpage

## Experimental Results by Individual News Sources {-}

```{r onlinepredictedbysource, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=10, out.width = "85%", fig.align = "center", fig.showtext=T, fig.cap = "Probability of evaluating news stories as true when they are attributed to specific state-run and independent media outlets, by approval of Vladimir Putin. Calculations based on a linear regression of news story evaluations on media outlet dummies and presidential approval (see text for details); results from the main study. 95\\% confidence intervals are shown"}

source_effect <- read_csv("Estimates/Figure_B5_main_treatment_by_source_estimates.csv") %>%
  mutate(pres_approval_cat = factor(pres_approval_cat, 
                    levels = c("Strong\ncritic",
                               "Moderate\ncritic",
                               "Moderate\nsupporter",
                               "Strong\nsupporter")),
         story_source = factor(story_source, 
                                      levels = c("Channel One", "Russia-24", 
                                                 "RIA", "RT", "KP",
                                                 "Meduza", "Rain", 
                                                 "Echo of Moscow")))

ggplot(source_effect, 
       aes(pres_approval_cat, estimate,
           color = story_source, shape = story_source)) +
  geom_point(size = 3,
             position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4)) +
  scale_color_manual(name = "Source", values = c("#a50026",
                                                 "#d73027",
                                                 "#f46d43",
                                                 "#fdae61",
                                                 "#fee090",
                                                 "#313695",
                                                 "#4575b4",
                                                 "#74add1")) +
  scale_shape_manual(name = "Source", values = c(17, 17, 17, 17, 17, 19, 19, 19)) +
  labs(x = "", y = "Probability of believing news stories, percent",
       title = "Perceived truthfulness of messages given news source") +
  ylim(c(35, 65)) +
  myggtheme +
  coord_flip()

```



\newpage

## Experimental Results for Pre-Selected and "Current" News Stories {-}

```{r onlinestateregimecurrent, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=8, out.width = "85%", fig.align = "center", fig.showtext=T, fig.cap = "The effect of changing the treatment from an independent to state media outlet on evaluations of news stories, by approval of Vladimir Putin and by story type. Calculations based on a linear regression of news story evaluations, accounting for state control and presidential approval; results from the main study. 95\\% confidence intervals are shown"}

cur <- read_csv("Estimates/Figure_B6_main_treatment_effect_current_estimates.csv") %>%
  mutate(pres_approval_cat = factor(pres_approval_cat, 
                    levels = c("Strong\ncritic",
                               "Moderate\ncritic",
                               "Moderate\nsupporter",
                               "Strong\nsupporter")))

ggplot(cur, 
       aes(pres_approval_cat, estimate,
           shape = story_current, linetype = story_current)) +
  geom_point(size = 3,
             position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 0.8) +
  geom_hline(yintercept = 0, linetype = "dashed", col = "gray") +
  labs(x = "", y = "Effect of state vs independent media",
       title = "The effect of state media by story type",
       linetype = "Stories", shape = "Stories") +
  ylim(c(-20, 20)) +
  myggtheme +
  coord_flip()


```


\newpage

## Experimental Results with Alternative Categorizations of State-Controlled Media Outlets {-}

In additional models, I consider alternative categorization of state-controlled media outlets. In the first model, _RBC_ is also a state-controlled media organization. (In the main analysis, _RBC_ is a separate category.) In the second model, I consider as state-controlled only those news outlets that are directly owned by the government: _Channel One_, _Russia-24_, _RIA_, and _RT_. _RBC_ and _KP_ are categorized as "Other." The results, reported in Figure \@ref(fig:onlinestateeffectalt) and in Table \@ref(tabonlinestatemediaeffectaltapproval) below, are very similar to the results in the main text.

```{r onlinestateeffectalt, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=8, out.width = "85%", fig.align = "center", fig.showtext=T, fig.cap = "Effect of changing the treatment from an independent to state media outlet on evaluations of news stories, by approval of Vladimir Putin. Here, RBC is considered as a state-controlled outlet. Calculations based on a linear regression of news story evaluations, accounting for state control and presidential approval; results from the main study. 95\\% confidence intervals are shown"}

alt_state_media <- read_csv("Estimates/Figure_B7_main_treatment_effect_alt_def_estimates.csv") %>%
  mutate(pres_approval_cat = factor(pres_approval_cat, 
                    levels = c("Strong\ncritic",
                               "Moderate\ncritic",
                               "Moderate\nsupporter",
                               "Strong\nsupporter"))) 

ggplot(alt_state_media, 
       aes(pres_approval_cat, estimate, linetype = state_definition, shape = state_definition)) +
  geom_point(position = position_dodge(width = 0.4), size = 3) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 0.8) +
  geom_hline(yintercept = 0, linetype = "dashed", col = "gray") +
  labs(x = "", y = "Effect of state vs independent media",
       title = "The effect of state media",
       subtitle = "For alternative definitions of state-controlled media",
       linetype = "", shape = "") +
  ylim(c(-20, 20)) +
  myggtheme +
  coord_flip()

```



\newpage

## Regression Tables for the Experiment {-}

\input{Tables/Table_B4_main_treatment_effect.tex}

\newpage

\input{Tables/Table_B5_main_treatment_by_source.tex}

\newpage

\input{Tables/Table_B6_main_treatment_effect_alt.tex}

\newpage

\input{Tables/Table_B7_main_treatment_effect_alt_def.tex}

\newpage

\input{Tables/Table_B8_main_treatment_effect_topic.tex}

\newpage

\setcounter{table}{0}
\renewcommand{\thetable}{C\arabic{table}} 
\setcounter{figure}{0} 
\renewcommand{\thefigure}{C\arabic{figure}}

\setcounter{page}{1}


# Online Appendix C: Additional Evidence From the Nationally Representative Survey (Study 2) {#appendixC -}

For practical reasons, the study on a nationally representative sample included three news stories from the online survey (two of them were shown in two versions; see below) and only two news sources, assigned randomly with an approximately equal probability: _Channel One_, the main state-run television station, and _Echo of Moscow_, a liberal radio station/website. Respondents saw the logo of either _Channel One_, or _Echo of Moscow_, and interviewers emphasized the name of the news organization before each news story. After each vignette, respondents were asked to evaluate the truthfulness of the message on a scale from 0 to 3 (rescaled in the analysis to take values from 0 to 1). 

The experimental vignettes and treatments were embedded in a nationally representative omnibus survey conducted monthly by a Russian polling firm, Levada Center. The omnibus survey uses in-home visits and relies on random sampling of the Russian population using a multi-stage sampling procedure (first randomly selecting urban and rural areas, then randomly selecting sampling stations within these primary sampling units, then randomly selecting households and individuals within households). The sample is stratified by sociodemographic characteristics based on the recent census data and on the recent demographic statistics, and weights are provided to further adjust for the discrepancies between the sample and the Russian population. The survey was fielded on August 22--28, 2019, covering 140 cities, towns, and rural settlements in 50 Russian regions. The sample size is 1608 respondents.

## News Stories in the National Survey {-}

**Economic struggles, version 1 (the Russian statistical agency, Rosstat, is not mentioned)**. _For 80% of Russian families, it is difficult to buy all the necessary goods and "make ends meet". More than half of the families cannot replace the simplest furniture that falls into disrepair._

**Economic struggles, version 2 (Rosstat is mentioned)**. _For 80% of Russian families, it is difficult to buy all the necessary goods and "make ends meet." This is what new research by the Federal service of government statistics says. More than half of the families cannot replace the simplest furniture that falls into disrepair._ (This version implies that the government has admitted the problem.)

**Ukrainian economy, version 1 (Russia is not mentioned)**. _The Ukrainian economy is growing at a slower rate than the world economy. According to analysts, in 2019, the world's GDP will grow by almost 4 percent, and the Ukrainian GDP by less than 3 percent._

**Ukrainian economy, version 2 (Russia is mentioned)**. _The Ukrainian economy is growing at a slower rate than the world economy, but faster than the Russian economy. According to analysts, in 2019 the world's GDP will grow by almost 4 percent, Ukrainian GDP by less than 3 percent, and Russian GDP by only 1.6 percent. The Ukrainian economy has been growing faster than the Russian economy for the fourth year in a row._ (This version is more politicized by including a direct comparison with Russia.)

**U.S. submarine**. _The U.S. submarine Hartford froze into Arctic ice during military exercises. The submarine was supposed to rehearse a Tomahawk launch against a hypothetical aggressor---Russian ships. But something went wrong, and the submarine could not rise to the surface. A helicopter had to be called in order to save the vessel from the captivity of ice._ (This is a fake story fabricated by the Russian state propaganda.)

## The Effect of State-Run Media, by Putin Approval {-}

Figure \@ref(fig:levadaeffectvote) shows the estimated effect of changing the treatment from _Echo of Moscow_ to _Channel One_. In the left panel, regime support is measured as respondent's vote choice in the last presidential election in order to account for the differences between different groups of Putin critics: liberal and pro-Western individuals, who are more likely to see the liberal-leaning _Echo of Moscow_ as like-minded, and nationalists or communists. In the right panel, regime support is measured as approval of Vladimir Putin. Also see Table \@ref(tablevadastatemediaeffect).

```{r levadaeffectvote, echo = FALSE, fig.show = "hold", fig.width=16, fig.height=8, out.width = "85%", fig.align = "center", fig.showtext=T, fig.cap = "The effect of changing the treatment from the independent (Echo of Moscow) to state-run (Channel One) media outlet on evaluations of news stories, by respondent's vote in the 2018 presidential election or by approval of Vladimir Putin. Calculations based on a linear regression of news story evaluations, accounting for state control, 2018 vote/Putin approval, and demographic covariates (see text for details); results from the national survey (Study 2). 95\\% confidence intervals are shown"}

lev1 <- read_csv("Estimates/Figure_C1_levada_treatment_effect_vote_estimates.csv") 

  lev2 <- read_csv("Estimates/Figure_C1_levada_treatment_effect_estimates.csv") %>%
  mutate(pres_approval_cat = factor(pres_approval_cat, 
                    levels = c("Strong\ncritic",
                               "Moderate\ncritic",
                               "Moderate\nsupporter",
                               "Strong\nsupporter"))) 

l1 <- ggplot(lev1, 
       aes(vote_outcome_2018, estimate,
           color = highlight)) +
  geom_point(size = 3,
             position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 0.8) +
  scale_color_manual(values = c("black", "gray")) +
  geom_hline(yintercept = 0, linetype = "dashed", col = "gray") +
  labs(x = "", y = "Effect of state media",
       title = "Given vote choice in 2018") +
  ylim(c(-27, 27)) +
  myggtheme +
  theme(legend.position = "none") +
  coord_flip()

l2 <- ggplot(lev2, 
       aes(pres_approval_cat, estimate)) +
  geom_point(size = 3,
             position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 0.8) +
  geom_hline(yintercept = 0, linetype = "dashed", col = "gray") +
  labs(x = "", y = "Effect of state media",
       title = "Given Putin approval") +
  ylim(c(-27, 27)) +
  myggtheme +
  coord_flip()

l1 + l2 +
plot_annotation(title = "The effect of state media in the national sample", theme = theme(plot.title = element_text(size = 30,
                                    family = "sourcesanspro")))

```

\newpage


## Regression Table for the Experiment {-}

\input{Tables/Table_C1_levada_treatment_effect.tex}

\newpage

\setcounter{table}{0}
\renewcommand{\thetable}{D\arabic{table}} 
\setcounter{figure}{0} 
\renewcommand{\thefigure}{D\arabic{figure}}

\setcounter{page}{1}

# Online Appendix D: Additional Evidence From the OMI Online Panel (the Media Perceptions Survey, Study 3) {#appendixD -}


## Questions About Individual News Sources {-}

[_These questions were asked for the following news outlets: RT, Channel One, Russia-24, RIA_]

**Would you say that these outlets provide a full sense of what is happening, do not ignore important topics or facts?** 

Mostly yes; Often ignore something important; Do not know the outlet well/difficult to say

**Would you say that these outlets provide accurate and truthful information?**

Mostly yes; Often provide false or inaccurate information; Do not know the outlet well/difficult to say

**Would you say that these outlets are politically unbiased, convey information in a neutral fashion?**

Mostly yes; Mostly convey information from the standpoint of the authorities; Mostly criticize the authorities; Do not know the outlet well/difficult to say

**Would you say that these outlets are independent in their editorial policies, they themselves decide what and how to cover?**

Mostly yes; The authorities decide for them; Do not know the outlet well/difficult to say


\newpage

## Regression Tables for the Media Perceptions Survey {-}

\input{Tables/Table_D1_media_usage_3_surveys.tex}

\input{Tables/Table_D2_OMI_media_trust.tex}

\newpage

\input{Tables/Table_D3_OMI_ranking_completeness.tex}

\input{Tables/Table_D4_OMI_ranking_accuracy.tex}

\newpage

\input{Tables/Table_D5_OMI_ranking_independence.tex}

\input{Tables/Table_D6_OMI_ranking_bias.tex}


\newpage

## Media Usage {-}

In all three surveys, I asked respondents to report the media outlets that they typically use to learn the news, and then I constructed dummy variables that indicate whether a respondent uses any of state-run television stations or any of critical news outlets. Then, I regressed these dummies on presidential approval and covariates, using the same model setup as with the analysis of media trust. Figure \@ref(fig:mediausage) plots the probabilities of using state-run television and foreign or critical media outlets across three samples.^[In the Levada survey, the definition of critical media is somewhat different: instead of naming specific news outlets, respondents indicated the usage of online/cable television channels (_Rain_ and _RBC_), business news outlets (most of which are editorially independent), and foreign websites. Combining these three categories, we can obtain an approximation for the usage of critical media, which, however, somewhat overstates it, as _RBC_ and some other business news outlets are influenced by the government.] Also see Table \@ref(tabmediausage).

```{r mediausage, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=8, out.width = "85%", fig.align = "center", fig.showtext=T,  fig.cap = "The probability of using independent media and state television, by approval of Vladimir Putin. Calculation based on linear regressions of media usage (dummy variables) on presidential approval and demographic covariates; results from the main study, from the nationally representative sample (Study 2), and from the OMI online panel (Study 3). 95\\% confidence intervals are shown"}

usage <- read_csv("Estimates/Figure_D1_media_usage_3_surveys.csv") %>%
  mutate(x = factor(x, levels = c("Strong\ncritic",
                               "Moderate\ncritic",
                               "Moderate\nsupporter",
                               "Strong\nsupporter")))

ggplot(usage, 
       aes(x, estimate, linetype = survey, shape = survey)) +
  geom_point(size = 3,
             position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4),
                size = 0.8) +
  labs(x = "", y = "Probability of using state TV or independent media",
       title = "Usage of state and independent media",
       linetype = "Sample", shape = "Sample") +
  ylim(c(0, 1)) +
  coord_flip() +
  myggtheme +
  facet_wrap(~media, ncol = 2)
```



\newpage

## Knowledge of Independent Media and Trust in/Usage of State Media {-}

Figure \@ref(fig:omitrustknowledge) shows the predicted probabilities of trust in state television and the usage of state television among supporters depending on whether they know of any critical news outlets or not (data from the OMI survey). The model builds on Figures \@ref(fig:omitrust) and \@ref(fig:mediausage), adding an interaction between approval and knowledge of independent media. Strong supporters trust state television a great deal regardless of their awareness of independent outlets. Moderate supporters who are aware of independent media may trust state television somewhat less, although the confidence intervals for two estimates overlap. The usage of state television similarly does not depend much on the knowledge of independent media.

```{r omitrustknowledge, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=7, out.width = "85%", fig.align = "center", fig.showtext=T,  fig.cap = "Probability of trusting or using state media depending on knowledge of independent media. Calculation based on a linear regression of media trust or media usage (dummy variables) on presidential approval, knowledge of independent media, and demographic covariates; results from the OMI online panel (Study 3). 95\\% confidence intervals are shown"}

omi_state_tv_aw <- read_csv("Estimates/Figure_D2_omi_state_tv_by_awareness.csv") %>%
  mutate(x = factor(x, levels = c("Moderate\nsupporter",
                               "Strong\nsupporter")),
         knowledge_indep = factor(knowledge_indep, 
                                  levels = c("Yes", 
                                             "No")))
                                             

ggplot(omi_state_tv_aw, 
       aes(x, estimate, linetype = knowledge_indep, 
           shape = knowledge_indep)) +
  geom_point(size = 3,
             position = position_dodge(width = 0.4)) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.4), size = 0.8) +
  labs(x = "", y = "Probability of trusting or using state TV",
       title = "Trust in/usage of state TV among supporters",
       subtitle = "Given knowledge of independent media",
       linetype = "Knows some independent media", 
       shape = "Knows some independent media") +
  ylim(c(0, 1)) +
  myggtheme +
  coord_flip() +
  facet_wrap(~outcome, ncol = 2)

```

\newpage

## Knowledge of Independent Media and the Evaluations of State Media {-}

The models here are analogous to the analysis of perceptions of accuracy and media bias in the main text; in this case, I add an interaction between approval and knowledge of independent media and control for the knowledge of the state media outlet in question.  

```{r omiperceptionsknowledge, echo = FALSE, fig.show = "hold", fig.width=14, fig.height=10, out.width = "85%", fig.align = "center", fig.showtext=T,  fig.cap = "Probability that Putin supporters evaluate state media negatively along various dimensions. Calculations based on multinomial regressions of news source evaluations on knowledge of independent media and covariates (see text for details); results from the OMI online panel (Study 3). 95\\% confidence intervals are shown"}

perc_supporters <- read_csv("Estimates/Figure_D3_omi_perceptions_by_awareness_supporters.csv") %>%
  mutate(source_known_independent = factor(source_known_independent,
                                                 levels = c("Yes", "No")),
                                                 criterion = factor(criterion,
                                   levels = c("Not independent",
                                              "Biased in favor of government",
                                              "Omits important facts",
                                              "Often gives false info")),
                                              source = factor(source, levels = c("Channel One", "RIA", "RT", "Russia-24")))

ggplot(perc_supporters, 
       aes(criterion, prob, 
           linetype = source_known_independent,
           shape = source_known_independent)) +
  geom_point(position = position_dodge(width = 0.6), size = 3) +
  geom_errorbar(aes(ymin = conf.low,
                    ymax = conf.high),
                width = 0,
                position = position_dodge(width = 0.6), size = 0.8) +
  labs(x = "", y = "Probability of agreeing with statements",
       title = "Perceptions of state media among Putin supporters",
       linetype = "Knows some independent media", 
       shape = "Knows some independent media") +
  ylim(c(0, 1)) +
  coord_flip() +
  facet_wrap(~source, ncol = 2) +
  myggtheme

```

